Built a CNN-based skin lesion classification model with convolutional filters, trained using RMSprop optimizer and KL Divergence loss function, achieving 77% accuracy on the HAM10000 dataset. Implemented data preprocessing techniques, including resizing images to 64x64 pixels, normalization, and balancing class distribution.

Utilized LSTM networks to fill temporal gaps in satellite NO2 data, leveraging its capability to handle time-series data and employed Random Forest Regression model for Spatial Downscaling. Integrated SRGAN for image smoothening, improving the visual quality of the output maps.

Built a RAG pipeline using Gemini-2.0-Flash and ChromaDB, with dynamic chunking and semantic retrieval, en abling context-aware responses based on the document. Integrated OCR using Tesseract to process scanned PDFs and retrieved the top-3 relevant chunks per query.